HAL Id: tel-01914986
https://pastel.archives-ouvertes.fr/tel-01914986
Submitted on 7 Nov 2018
HAL is a multi-disciplinary open access
archive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés.
transportation problems
Luca Alberto Davide Ferrari
To cite this version:
Luca Alberto Davide Ferrari. Phase-field approximation for some branched transportation problems. Optimization and Control [math.OC]. Université Paris Saclay (COmUE), 2018. English. �NNT : 2018SACLX049�. �tel-01914986�
NNT
:2018SA
CLX049
champs de phases pour des
problèmes de transport branché
Thèse de doctorat de l’Université Paris-Saclaypréparée à l’Ecole Polytechnique Ecole doctorale n◦573 : Interfaces (approches
interdisciplinaires/fondements, applications et innovation)
Spécialité de doctorat : Mathématiques appliquées
Thèse présentée et soutenue à Palaiseau, le 5/10/2018, par
L
UCAA
LBERTOD
AVIDEF
ERRARIComposition du Jury :
Filippo Santambrogio
Professeur, Université Paris Sud Président Ilaria Fragalá
Professeur, Politecnico di Milano Rapporteur Antoine Lemenant
Maître de Conférences, Université Paris Diderot Rapporteur Antonin Chambolle
Directeur de recherche, École Polytechnique Directeur de thèse Benoit Merlet
Professeur, Université Lille 1 Co-directeur de thèse Lucia Scardia
Professeur, University of Bath Examinateur Blanche Buet
Maître de Conférences, Université Paris Sud Examinateur Edouard Oudet
First of all, I would like to express my deep gratitude to my two supervisors, Antonin Chambolle and Benoit Merlet, for their guidance. I will always be grateful for the suggestions on the subjects and I am grateful to Prof. Chambolle for his inspirations. I will always remember the first time he introduced to me the theory of Gamma con-vergence and the subject of branched transportation. Seeing him work at a blackboard is something really fascinating.
I am grateful to Prof. Merlet: he introduced me to the complex and fascinating world of research. I learnt from him how to approach a problem, how to write a paper and how to present my results. He was always so kind to go through my writings and suggest how to present my works in a better way and he was always pleased to get his hands dirty and discuss on the most technical difficulties of a proof.
I owe my heartfelt gratitude to Benedikt Wirth e Carolin Rossmanith for their welcome in Munster and for the passionate exchange of ideas that lead to a very nice work written together.
I would like to thank Prof. Lemenant and Prof. Fragala for agreeing to review the current thesis and their suggestions and corrections. I am also grateful to Filippo Santambrogio, Edouard Oudet, Blanche Bouet and Lucia Scardia for accepting to be examinateurs at my Ph.D. defense.
Let me thank all the collegues, Adi, Marco, Valentina, Belhal, Remi, Genevieve, Simona, Matteo, Luca, Vito and Fred at CMAP that have accompanied me during these years of research.
Grazie alle mie due famiglie adottive a Parigi Tommaso, Marta e Gregorio e Mar-tino, Federica e Mattia per tutte quelle volte che vi siete sentiti dire che non avrei mai finito questa tesi ed invece eccoci qui a discuterla.
Un grazie immenso a tutti i miei compagni d’appa passati e presenti Francesco, Tommaso, Matteo, Emanuele, Martino, Giacomo, Pietro, Matteo, Filippo e Pietro. Perch tornare a casa e trovare degli amici con cui si condivide la vita una cosa grandiosa che mi mancher, come mi mancheranno i cartelli stradali appesi a Place d’It e le poltrone trovate per strada.
Un grazie gigante a tutti i miei amici parigini a quelli che son venuti e poi partiti e a quelli che son ancora qui. Hlose, Marta, Lucia, Caterina, Clelia, Paince, Roby,
Grazie a tutti i miei amici ”italiani” Maddalena, Pietro, Elisa, Silvia, Francesca, Elisa che pur da lontano mi hanno fatto compagnia.
Un grazie di cuore alla mia famiglia in particolare mia madre e mio padre per tutto il loro sostegno e per la passione al lavoro che mi hanno insegnato semplicemente mostrandomi il loro modo di lavorare.
L’ultima immensa gratitudine verso Chiara, che in questi tre anni e mezzo di distanza ha comunque trovato il modo di essere sempre presente nel mio cuore e ac-compagnarmi nella mia strada anche se distante da lei.
Introduction 1
1 Affine cost function 17
1.1 Introduction . . . 17
1.2 Preliminaries for the chapter . . . 19
1.3 Local Result . . . 20
1.4 Equicoercivity and Γ-liminf . . . 26
1.5 Γ-limsup inequality . . . 28
1.6 Numerical Approximation . . . 33
2 Multidimensional case 39 2.1 Introduction . . . 39
2.2 Reduced problem results in dimension n− k . . . . 41
2.3 Compactness . . . 42
2.4 Γ-liminf inequality . . . 47
2.5 Γ-limsup inequality . . . 50
3 The k-dimensional problem 55 3.1 Introduction . . . 55
3.2 Compactness and k-rectifiability . . . 56
3.3 Γ-liminf inequality . . . 60
3.4 Γ-limsup inequality . . . 61
3.5 Discussion about the results . . . 63
4 Piecewise affine cost functions 65 4.1 Introduction . . . 65
4.2 Remarks . . . 67
4.3 The Γ-limit of the phase field functional . . . 69
4.3.1 The Γ− lim inf inequality for the dimension-reduced problem . 69 4.3.2 The Γ− lim inf inequality . . . . 76
4.3.3 Equicoercivity . . . 78
4.3.4 The Γ− lim sup inequality . . . . 79
4.4 Numerical experiments . . . 85
4.4.1 Discretization . . . 85
4.4.2 Optimization . . . 85
5 Generalized cost functions 95
5.1 Introduction . . . 95
5.2 Origin of the model and preliminaries . . . 96
5.3 Proof of Theorem 5.1 . . . 101
Conclusion 107 A Density result for vector measures in R2 111 B Reduced problem in dimension n− k 115 B.1 Auxiliary problem . . . 115
B.2 Study of the transition energy . . . 118
B.3 Proof of Proposition 2.1 . . . 121
B.4 Proof of Proposition 2.2 . . . 123
B.5 Proof of Proposition 2.3 . . . 124
C Slicing of measures 127
When designing a supply-demand distribution network it is convenient to give it a tree structure in which it is preferable to regroup mass in the transportation process. This assumption emerges from numerous observations, for instance the structure of the blood vessels in the cardiovascular system is required to distribute blood from a concentrated source in the heart to a widespread volume or vice-versa, the root system of a tree needs to recollect water from the soil. In these situations we may observe how broad and long vessels are preferable rather than thin spread out ones. The assumption we make is that the actual observed network is optimal with respect to some given cost among all possible networks developing from a source and irrigating a given sink. These structures appear in a wide range of situations (figure 1) and many efforts have been made by the mathematical community in order to give a precise model able to describe all the observable features of these networks.
Figure 1: On the left: root network of a tree. On the right: angiography of an eye in which it is possible to recognize the tree structure of the network of blood vessels.
A first well known approach in the framework of graph theory was proposed by Gilbert in [PST15] where he deals with the Steiner Minimal Tree [AT04, PS13] problem. The latter consists in finding the graph connecting a given set of points {x0, . . . , xN}
with minimal total length. More formally, a Steiner minimal tree is the solution of the variational problem
argminH1(K) : K compact, connected and contains x0, . . . , xN , (1)
where H1(K) is the Hausdorff 1-dimensional measure of F (the length of K, if it is 1-dimensional and sufficiently smooth). As stated in Courant and Robbins [CR79] the Steiner minimal tree problem can be thought of as a naif model for the network of highways connecting a set of cities. The drawback of the model is that the local intensity of the traffic is not taken into account. Nevertheless it allows to appreciate the issues emerging from these models. As observed in the quoted paper [PST15] in a Steiner minimal tree, differently from the Minimal Spanning Tree [Kru56], new vertices
Figure 2: Steiner Minimal Tree connecting 10000 points randomly distributed in the plane. The problem was solved using the GeoSteiner algorithm [WZ97], which is cur-rently the most efficient exact algorithm for computing minimum Steiner trees.
may be added in order to minimize the total length thus, rather than the network itself, the real unknown is its topology. An example of this situation is shown in Figure 3. This feature appears as well in other models in which the cost per unit length depends on the intensity of the traffic flux [Gil67]. In light of this high combinatorial complexity, the problem is in the list of NP-complete problems from Karp [Kar72] and it is still an active field of research even in the operational research community [FMBM16].
(1, 0) (−1/2,√3/2) (−1/2, −√3/2) (1, 0) (−1/2,√3/2) (−1/2, −√3/2)
Figure 3: On the left: Minimal Spanning Tree connecting three points situated at the vertices of an equilateral triangle (length = 2√3). On the right: Steiner Minimal Tree constrained to connect the same set of points (length = 3). In dark blue the additional vertex which allows to decrease the total length.
The purpose of this thesis is to devise approximations of some Branched Transporta-tion problems. Branched TransportaTransporta-tion is a mathematical framework for modeling supply-demand distribution networks which is more general than the Steiner prob-lem presented above. In particular the supply factories and the demand locations are
modeled as measures supported on points and the network is interpreted as a vector measure, eventally the problem is cast as a constrained optimization problem. Given a function h, the transport cost of a mass m along an edge with length ` is h(m) ` and the total cost of a network is defined as the sum of the contributions of all its edges. The branched transportation case corresponds to the specific choice h(m) =|m|α with α ∈ [0, 1). The sub-additivity of the cost function, h(m1 + m2) ≤ h(m1) + h(m2),
ensures that transporting two masses jointly is cheaper than doing it separately. This formulation shares much of the numerical complexities presented above in the case of the Steiner Minimal tree problem. In this work we introduce various variational ap-proximations by means of elliptic-type functionals to obtain more efficient numerical schemes. Eventually the proposed method is generalized to Plateau-type problems, which is a framework to model soap films spanning a given boundary. In its more general formulation the unknown of this problems is a k-dimensional surface in Rn
spanning a (k − 1)-dimensional boundary and minimizing a certain cost. Branched transportation corresponds to a Plateau type problem for the choice k = 1.
↑
↑ ↑
Figure 4: Example of a surface spanning a 1-dimensional boundary consisting of three oriented circles.
Description of the model
Let us introduce precisely the framework for Branched Transportation [BCM09, Vil03]. First we introduce transport networks in a open set Ω∈ Rn, and the associated cost functional. For this purpose consider a segment Σ⊂ Ω, a positive real number m ∈ R+
and the vector τ ∈ Sn−1 tangent to Σ. The writing
m τH1
x
Σ (2)defines a vector valued measure, whereH1
x
Σ is the Hausdorff 1-dimensional measurein Rn restricted to the segment Σ. Intuitively, the Radon measure H1
x
Σ associates to any measurable set A the length of A∩ Σ. We say that a vector valued measure σ ∈ M(Ω, Rn) is polyhedral if it is a finite sum of measures of the form (2), namelyσ =P
imiτiH1
x
Σi. The action of σ on C0(Ω, Rn) is defined by the formula(σ, ϕ) =X
i
Z
Σi
A transport cost function h : R→ [0, +∞) is an application such that h is
(
even, lower semicontinuous,
sub-additive, with h(0) = 0. (3)
Given a transport cost function h we define the Gilbert energy of a polyhedral vector measure σ as
Eh(σ) :=
X
i
h(mi)H1(Σi).
We endow M(Ω, Rn) with its weak-∗ topology and extend E
h by relaxation, namely
for a vector measure σ ∈ M(Ω, Rn) we let Eh(σ) := inf lim inf j→+∞ Eh(σj) : σj ∗ * σ and σj polyhedral . (4)
By White in [Whi99a, 6] conditions (3) are sufficient in order to extendEhonM(Ω, Rn).
Choosing h(m) = |m| in equation (4) we obtain the mass functional which associates to each vector measure σ its total variation
|σ| = sup{(ϕ, σ) : ϕ ∈ C0(Ω, Rn), kϕk∞ ≤ 1}.
Otherwise, with h(m) = χ{m6=0}, where χ denotes the characteristic function of a set,
Eh reduces to the size functional which measures the length of the support of σ, namely
σ 7→ H1(supp(σ)). Other remarkable choices are represented in Figure 5.
1) m h(m) h(m) =|m| 2) m h(m) h(m) =|m|α 3) m h(m) h(m) = χ{m6=0} 4) m h(m) h(m) = (1 + β|m|)χ{m6=0} 5) m h(m) h(m) = min{α0|m|, α1|m| + β1}
Figure 5: For h as in the graphs we obtain respectively the: 1) Mass, 2) α-Mass, 3) Size, 4) Affine cost, 5) Urban planning functional.
To model the source and the sink of the transport network we introduce two probability measures µ+, µ− ∈ P(Ω) and restrict our attention to the vector space
Xµ+,µ− ⊂ M(Ω, Rn) consisting of those vector measures σ satisfying
div σ = µ+− µ− (5)
in the sense of distributions. As shown in the note [CFM18] if the relaxation is obtained with respect to polyhedral measures in Xµ+,µ− we still obtain the functional (4).
Finally we are interested in approximating minimizers of the Gilbert energy under the divergence constraint (5), namely:
min{Eh(σ) : σ ∈ Xµ+,µ−} . (6)
The Branched Transportation case corresponds to the choice h(m) = |m|α with α ∈
[0, 1) and has been introduced by Xia who has investigated as well the problem of existence and regularity of solutions. In [Xia03] the author, taking advantage of vari-ationals methods, proves the following
Theorem 0.1 (Existence Theorem). Given α∈ (1−n1, 1] and two probability measures µ+, µ− ∈ P(Ω), there exists a vector valued measure σ ∈ Xµ+,µ− for which Eh(σ) is
minimal. Furthermore for ˆσ ∈ argmin Eh(σ) we have the following estimate
Eh(ˆσ) ≤ 1 21−n(1−α)−1 √ n diam(Ω) 2 .
In a subsequent result [Xia04, Theorem 2.7] the same author addresses the problem of regularity. To state the result we need to introduce the notion of rectifiable vector measure. Namely a vector measure σ is said rectifiable if
σ = m τH1
x
Σ (7)where Σ, the support of σ as a distribution, is an H1-rectifiable set, its H1-density is the function m ∈ L1(H1
x
Σ) and τ : Σ → Sn−1 spans for H1-a.e. point in Σ the tangent space to Σ. In the following we denote with (m, τ, Σ) the rectifiable measure σ defined in (7).Theorem 0.2 (Structure of finite energy networks). Let µ+, µ− ∈ P(Ω). For 0 ≤ α <
1 if σ ∈ Xµ+,µ− is of finite total variation and finite E
h energy then it is rectifiable.
Furthermore if σ = (m, τ, Σ) we have
Eh(σ) =
Z
Σ
|m|α dH1. (8)
Equation (8) is of particular relevance since it extends the explicit representation of the functional to any rectifiable measure. The case of general vector valued measures and general transport cost functions h has been taken into consideration by Brancolini and Wirth in [BW18, Proposition 2.32] which shows that
Proposition 0.1 (Generalized Gilbert-Steiner Energy). Let µ+, µ− ∈ P(Ω), σ ∈
M(Ω, Rn) with finite total variation and such that div σ = µ
+− µ− then σ can be
decomposed as
σ = σ⊥+ m τH1
x
Σwhere (m, τ, Σ) is the H1-rectifiable component of σ and σ⊥ is the diffused one.
Fur-thermore
Eh(σ) = h0(0)|σ⊥|(Ω) +
Z
Σ
h(m) dH1. (9) With an abuse of notation we have denoted h0(0) = limm↓0h(m)/m.
Before introducing problems involving surfaces and other higher dimensional objects let us highlight the fact that the Steiner minimal tree problem connecting some points {x0, . . . , xN} may be modeled in the context of Branched transportation. Firstly, with
the choice α = 0, Eh reduces to the size functional. Secondly the divergence constraint
forces any considered vector measure to join the support of µ+ to the support of µ−
thus, by choosing µ+= δx0 and µ− = 1/N
PN
i=1δxi we force x0 to be connected to each
xi. Gathering all together, with these choices, a minimizer σ of (6) is supported on
a set connecting the points in {x1, . . . , xN} to x0 and has support with minimal total
length thus is a solution to (1).
The energy introduced above for rectifiable measures supported on 1-dimensional surfaces can be generalized to any dimension k ∈ {1, . . . , n}. To this aim is necessary to introduce the concept of k-currents in Rn. Denote with Dk(Ω) the space of smooth
differential forms on the open set Ω. The vector space of k-currents, Dk(Ω), is the dual
to Dk(Ω) and it is naturally endowed with its weak-∗ topology. We mainly follow the
notation of [KP08, Fed69] the main difference being the use of σ to denote a k-current instead of a latin capital letter. For a current we can define a notion of boundary by duality as follows
h∂σ, ωi = hσ, dωi for all (k− 1)-differential forms ω.
We call mass of a k-current the supremum of hσ, ωi among all k-differential forms with comass bounded by 1, and denote it with |σ|. In particular by the Radon-Nikodym theorem we can identify a k-current σ with finite mass with the vector valued measure τ µσ where µσ is a finite positive valued measure and τ is a µσ-measurable map in the
set of unitary k-vectors for the mass norm. The relation with vector measure is evident when we consider the fact that the vector spaces Λ1Rn, Λ1Rn identify with Rn. Thus
any vector measure σ ∈ M(Ω, Rn) with finite mass identifies with a 1-current with finite mass and vice-versa. Furthermore the divergence operator acting on measures in the sense of distributions is defined by duality as the boundary operator for currents. Thus, in analogy with what has been presented for vector measures, in equation (7), a k-current σ is said to be k-rectifiable if we can associate to it a triplet (θ, τ, Σ) such that
hσ, ωi = Z
Σ
θhω, τi dHk
where Σ is a countably k-rectifiable subset of Ω, τ at Hk a.e. point is a unit simple
values in R+. The vector space of Rectifiable Currents is denoted with Rk(Ω). Among
these we single out the subset Pk(Ω) of rectifiable currents for which Σ is a finite union
of polyhedra and θ is constant on each of them, these will be called Polyhedral Chains. For any k-current σ such that both σ and ∂σ are of finite mass we say that σ is a normal k-current and we write σ∈ Nk(Ω). On the space Dk(Ω) we can define the flat
norm by
F(σ) = inf{|σR| + |σS| : σ = σR+ ∂σS where σS ∈ Dk+1(Ω) and σR ∈ Dk(Ω)} ,
which metrizes the weak-∗ topology on currents on compact subsets of Nk(Ω). Finally
the flat chains Fk(Ω) consist of the closure of Pk(Ω) in the F topology. By the scheme
of Federer [Fed69, 4.1.24] it holds
Pk(Ω) ⊂ Nk(Ω) ⊂ Fk(Ω).
Following the strategy proposed by Fleming [FF60, Fle66] in the context of flat chains with coefficients in groups we now define the energy Eh on the space of flat
chains. Let h be a transport cost function and σ =P(miτi, Σi) a polyhedral current
we let
Eh(σ) :=
X
i
h(mi)Hk(Σi).
In analogy to what has been done before we extend Eh on the space of flat chains by
relaxation. For σ∈ Fk(Ω), Eh(σ) := inf lim inf j→∞ Eh(σj) : σj polyhedral and F(σj − σ) → 0 . (10)
In Chapter 3 we look for approximations to problems of the type
min{Eh(σ) : ∂σ = ∂σ0} (11)
where σ0 is a given polyhedral k-current. These problems have been introduced and
studied in [Mor89, DPH03] by Morgan, De Pauw and Hardt among others to propose different models for soap film minimal surfaces. The latter is the k-dimensional gener-alization to the minimization problem defined in (6). As sketched in [Whi99a, Whi99b] Eh has an explicit formulation on rectifiable currents, namely for a rectifiable current
(m, τ, Σ) we have
Eh(σ) :=
Z
Σ
h(m) dHk.
This result has been proved in [CDRMS17, Proposition 2.6], is and is the consequence of the following polyhedral approximation theorem
Theorem 0.3 (Polyhedral approximation). Let h be a transport cost function and let σ = (m, τ, Σ) be a rectifiable k-current. For every δ > 0 there exists a polyhedral k-chain ˆσ =P(mi, τi, Σi) such that
F(ˆσ− σ) ≤ δ, X i h(mi)Hk(Σi)≤ Z Σ h(m) dHk+ δ and |ˆσ| ≤ |σ| + δ.
In addition Colombo et al. in [CDRMS17, Proposition 2.7] have shown that the condition lim m↓0 h(m) m = +∞ is equivalent to the fact
Eh(σ) finite if and only if σ is rectifiable.
This result may be seen in correlation with equation (9) presented above. Let us highlight that the polyhedral approximation result from Colombo et al. does not take into account any boundary constraint for the k-currents. An analogous result with boundary constraint has been proved in the note [CFM18]. We conclude this section with an important sufficient condition for a flat chain to be rectifiable, proved by White in [Whi99a, Corollary 6.1].
Theorem 0.4 (Rectifiability for currents). Let σ ∈ Nk(Ω) be a normal k-current
supported on a k-rectifiable set; then σ is rectifiable.
We will take advantage of this theorem even in the context of vector measures. With the notation introduced above it reads as
Theorem 0.5 (Rectifiability for vector-valued measures). Let σ ∈ M(Ω, Rn). If |σ|(Ω) + |∇ · σ|(Ω) < ∞, ∇ · σ is at most a countable sum of Dirac masses and there exists a Borel set Σ with H1(Σ) <∞ and σ = σ
x
Σ, then σ is a rectifiable vectormeasure.
Variational approximation for minimization problems
We provide approximations to the problems defined in (6) in the sense of Γ-convergence. The latter is a notion of functional convergence introduced by De Giorgi [DG75] to deal with variational problems. Following [DM93, Bra98, AD00, Bra02] we give the operative definition of Γ-convergence.
Definition 1 (Γ-convergence). Let X be a metric space, and for ε > 0 let Fε : X →
[0, +∞]. We say that FεΓ-converges to F : X → [0, +∞] on X as ε → 0 and we note
Fε Γ
−
→ F if the following two conditions hold:
(LB) Γ− lim inf inequality: for any x ∈ X and any xε → x it holds
lim inf
ε→0 Fε(xε)≥ F (x),
(UB) Γ− lim sup inequality: for any x ∈ X there exists a sequence (ˆxε)⊂ X such that
ˆ
xε → x and
lim sup
ε→0 Fε
The sequence (ˆxε) si called recovery sequence for x. Condition (UB) is frequently
hard to prove thus it is convenient to find a subset D ⊂ X such that: for every x ∈ X there exists an approximating sequence (xn) ⊂ D such that xn → x and
F (xn) → F (x). If we are able to recover D then a simple diagonal argument shows
that it is enough to verify condition (UB) for all x∈ D rather than for every x ∈ X. In the context of our work the set D corresponds with the set vector space of polyhedral vector measures. Since the definition of Γ-convergence may appear cumbersome let us provide this alternative characterization that allows to appreciate its relevance in the context of the Calculus of Variations.
Theorem 0.6 (Characterization of Γ-convergence). Let X be a metric space, and for ε > 0 let be givenFε : X → [0, +∞] and F : X → [0, +∞]. Fε
Γ
−
→ F if and only if for every G continuous functional, if xε minimizes Fε+G and xε → x then x minimizes
F + G .
Our strategy is to replace the singular energyEhwith a sequence of smoother elliptic
type functionals Fε and prove that Fε Γ
−
→ Eh. Then we prove that the family (Fε)
is equicoercive: any sequence of minima (ˆxj) is precompact in X. This ensures that
the sequence of minimizers ˆxε converge to a minimum. Finally we look for numerical
methods to approximate a minimum ˆxε.
Let us present three remarkable examples of Γ-convergence: Modica-Mortola, Am-brosio Tortorelli and a variation of the latter. Consider a container Ω⊂ R3 of unitary volume containing two immiscible liquids modeled by a binary function ϕ : Ω→ {0, 1} so that RΩ|ϕ| dx = V ∈ (0, 1) represents the percentage of one liquid with respect to the container’s volume. We associate to the system an energy depending on the surface tension, by supposing that it is directly proportional to the area of the interface Jϕ
between the liquids
M (ϕ) = cH2
(Jϕ). (12)
An alternative way to model this system is to assume that the transition is not given by an infinitesimal separating interface, but is rather a continuous phenomenon occurring in a thin layer of size ε. In view of this Cahn and Hilliard [CH58] consider a continuous phase function ϕ : Ω→ [0, 1] representing the pointwise mixing between the fluids and postulate an energy of the type
Z Ω ε2 |∇ϕ|2+ ϕ2(1 − ϕ)2 dx. (13)
The term ϕ2(1− ϕ)2 is called a double well potential and penalizes values far from 0 or 1; inhomogeneity is unfavoured by the gradient term. The link between (12) and (13) was discovered by Modica and Mortola in their papers [MM77a, MM77b]. Their result is more general, as a matter of fact, they prove that a suitable rescaling of the above energy Γ-converges to the perimeter functional in any domain dimension.
Theorem 0.7. Let Ω ⊂ Rn, and let X = BV (Ω)∩ L∞(Ω). For ϕ
∈ X, V > 0 and ε > 0, set Mε(ϕ) := Z Ω ε|∇ϕ|2+ ϕ 2(1− ϕ)2 ε dx, if ϕ∈ W1,2(Ω, [0, 1]) and Z Ω|ϕ| dx = V, +∞, otherwise in X.
Let c = 2R01t2(1− t)2 dt and M (ϕ) := ( cHn−1(Jϕ), if ϕ = χA and |A| = V, +∞, otherwise in X. Then Mε Γ − → M as ε → 0 in the L1 topology.
In the above BV(Ω) denotes the space of those functions ϕ such that ϕ∈ L1(Ω) and
the distributional gradient Dϕ is a Radon measure. For Bounded Variation functions the distributional gradient can be decomposed into three measures, namely
Dϕ =∇ϕ + Dcϕ + [ϕ]Hn−1
x
Jϕwhere∇ϕ is the component of Dϕ absolutely continuous with respect to the Lebesgue measure, Dcϕ is a Cantor measure and [ϕ]Hn−1
x
Jϕ is called the jump componentof the measure and is absolutely continuous with respect to the measure Hausdorff measure Hn−1 restricted to the discontinuity set Jϕ. In particular if ϕ ∈ BV(Ω) and
ϕ = χA then Jϕ is the essential boundary of A contained in Ω and [ϕ] = 1. For further
results on the theory of functions of Bounded Variations we refer to [AFP00] and the technical introduction of Chapter 1, Section 1.2. Theorem 0.7 is correlated with its respective equicoercivity property.
Corollary 0.1. If ε ↓ 0 and ϕε minimizes Mε then the sequence (ϕε) si pre-compact
with respect to the weak-* topology in BV and any limit point minimizes M .
Another example comes from the approximation of the Mumford-Shah functional for image segmentation. In [MS89] the authors consider a function g, defined on a domain Ω, representing the gray scale values of an image of a group of objects given by a camera, with discontinuities along the edges of the objects. The idea is that the segmented image u should be sufficiently smooth outside an (n− 1)-dimensional set containing the discontinuity set K, namely u ∈ W1,2(Ω\ K), and the latter should
be chosen of minimal Hn−1-size. Therefore they propose to optimize in the variables (u, K) the energy
Z
Ω\K
|∇u|2+ α(u
− g)2
dx + βHn−1(K).
The parameters α, β control the weight between the fidelity term|u−g|2 and the size of
the discontinuity set K. It is convenient to recast the problem in its weak formulation letting u∈ BV(Ω) and replacing the set K with Ju obtaining the functional
S (u) :=Z
Ω
|∇u|2+ α(u
− g)2
dx + βHn−1(Ju).
To give an approximation of the energyS , Ambrosio and Tortorelli have proposed the family of functionals Sε(u, ϕ) = Z Ω |∇u|2ϕ +β 4 ε|∇ϕ|2+ (1− ϕ) 2 ε dx + α Z Ω (u− g)2 dx.
In the articles [AT90, AT92] it is proved that Sε Γ
−
→ S . Let us give a heuristic idea behind this result. Since u is close to g, in the event of a strong discontinuity of g the gradient term |∇u| explodes. Indeed, high values in the gradient |∇u| are controlled by values close to zero in the state function ϕ. On the other hand the term in square brackets strongly penalizes values of ϕ far from 1. The competition of the terms in ϕ results in the fact that 1− ϕ represents a smoothed version of the function 1 − χJu.
Finally in the limit ε↓ 0 the Modica-Mortola term converges to the Hn−1 size of the
set{ϕ 6= 1} which contains the jump set of u. Functionals modeled on the ones from Ambrosio and Tortorelli and the latter functional itself are frequently known under the name of phase-field approximations. This is not only because of the strict relation with the Modica-Mortola functional but even because we may interpret the function ϕ as a state function, which acquires value 0 on the jump set of u, i.e. on the set of strong discontinuity of the function, and value 1 where u is sufficiently smooth. The two behaviors of u are then interpreted as two possible states and ϕ models the pointwise state function for the system. This observation has been taken into consideration in the work on fracture theory from Iurlano et al. [CFI16, Iur13]. There ϕ models the damage state of a material and u is replaced with a displacement function.
To conclude the section we present a variation of the Ambrosio Tortorelli functional proposed by Bonnivard, Lemenant and Santambrogio [LS14, BLS15] to recover in the limit the functional associated to the Steiner minimal tree problem for some points {x0, . . . , xN} ⊂ Ω ⊂ R2. Given a continuous function ϕ : Ω → [0, 1] the authors
introduce a geodesic distance weighted on ϕ, namely
dϕ(x, y) = inf Z γ ϕ dH1 : γ ∈ C([0, 1], Ω), γ(0) = x, γ(1) = y .
The distance dϕ(x, y) vanishes if and only if the two points x, y are joined by a path
on which ϕ is equal to 0. Now consider the functional
Z Ω ε|∇ϕ|2+(1− ϕ) 2 4ε dx + 1 cε N X i=1 dϕ(x0, xi)
where cε→ 0 as ε → 0. First observe that if N
X
i=1
dϕ(x0, xi) = 0 (14)
then the set{ϕ = 0} should include a path-connected subset containing {x0, . . . , xN}.
The heuristic argument for the Γ-convergence result follows the ideas presented in the case of the Ambrosio-Tortorelli functional. The exact result in [BLS15] is
Theorem 0.8 (Bonnivard-Lemenant-Santambrogio). Let Ω ⊂ R2 be an open set,
{x0, . . . , xN} ⊂ Ω and µ = N1
PN
i=1δxi. Consider the functional
Bε(ϕ) = Z Ω ε|∇ϕ|2+(1− ϕ) 2 4ε dx + Z Ω 1 cε dϕ(x0, x) dµ, (15)
and a sequence ϕε such that
Bε(ϕε)− inf
ϕ Bε(ϕ)−−→ε→0 0.
Then the sequence of functions dϕε converges uniformly (up to a subsequence) to a
function d such that the set K := {d = 0} minimizes H1 among all compact, connected sets containing the points {x0, . . . , xn}.
A first approach to the problem of approximating the energyEh in the case h =|·|α
was proposed by Santambrogio and Oudet in [OS11]. They introduce a functional of the type
Z
Ω
εα+1|∇σ|2+ εα−1|σ|β with σ ∈ W1,2(Ω, R2) and ∇ · σ = (µ+− µ−)∗ ρε
with β = (4α− 2)/(α + 1) and ρε an approximation to the identity. Actually the
com-plete Γ− lim sup inequality for the latter result has been provided by Monteil [Mon15, Mon17].
Structure of the thesis
In the First Chapter we study a variation of the functional proposed by Lemenant and Santambrogio. Motivated by the observation that
dϕ(x, y) = min
Z
Ω
ϕ|σ| dx : σ ∈ M(Ω, Rn) and div σ = δx− δy
we replace the term depending on the geodesic distance in (15) with a term depending on the product ϕ|σ|. The proposed functional is defined on couples (σ, ϕ) is
Z Ω ϕ|σ|2 2ε dx + Z Ω ε 2|∇ϕ| 2+(1− ϕ)2 2ε dx,
where σ is a vector-valued function complemented with the constraint
div σ = (µ+− µ−)∗ ρε. (16)
In the above ρε is a given approximation of the identity and the phase functions ϕ∈
L1(Ω) are bounded from below by the quantity β ε, where β ≥ 0 a given parameter. First, we show that this functional Γ-converges to the energy Eh, for the choice
h(m) := (
1 + βm, if m6= 0,
0, otherwise. (17)
The proof of the Γ-convergence result is obtained for open convex subsets of R2. The
advantage of choosing a quadratic penalization in σ is that the augmented Lagrangian problem associated to the functional may be explicitly solved in the dual variable. Therefore it is possible to devise an alternate minimization algorithm composed of two smooth elliptic functionals solvable via finite elements methods. The algorithm is
proposed and studied at the end of the chapter. We further present and study other algorithms which take advantage of a concept of ’shape derivative’ to improve the quality of the approximation.
The generalization to Ω⊂ Rnis the matter of the Second Chapter. To obtain the result in higher dimension the Modica-Mortola component of the functional needs to be rescaled. As observed in [Ghi14] this leads to the introduction of some non linearities in the functional as follows
Z Ω ϕ|σ|2 ε dx + Z Ω εp−n+1|∇ϕ|p+(1− ϕ) 2 εn−1 dx, (18)
for some p > n− 1. Again σ is complemented with the divergence constraint (16) for a suitable choice of ρε and we require a lower bound for the the phase field functions,
namely ϕ≥ βεn. We prove the Γ-convergence of the above functional to E
hn−1β where
the cost function hn−1β is the limit in ε of an optimization problem depending on the co-dimension n− 1. Namely for a ball Br ⊂ Rn−1 we let
hn−1ε,β (m) = min Z Br ϕ|θ|2 ε + ε p−n+1 |∇ϕ|p +(1− ϕ) 2 εn−1 dx, ϕ∈ W1,p(Br), ϕ = 1 on ∂Br and Z Br θ dx = m.
The latter optimization problem corresponds to the 0-dimensional version of (18). We introduce and study hdε,β (obtained replacing n− 1 with d in the latter formula) in the appendix. Some similar phase transition problems with mass constraint which leads to measures concentrated on atoms have been studied by Bouchitt´e, Dubs and Seppecher in [BDS96] in the context of droplets equilibrium. In particular we show that hdβ is independent of r and that it is a transport cost function satisfying the conditions (3). We prove as well that there exists a constant c > 0 such that
1 c ≤
hdβ(m)
1 +√βm ≤ c for m > 0.
Remark that the Modica-Mortola component of the functional studied in the second chapter depends on n− 1, the co-dimension of the problem in the case of rectifiable measures. In Chapter Three we investigate a different rescaling to approach minima to (11) defined for k-currents, namely
Z Ω ϕ|σ|2 ε dx + Z Ω εp−n+k|∇ϕ|p+ (1− ϕ) 2 εn−k dx. (19)
In this context, σ is no longer a vector measure, to take into account the boundary the constraint needs to be suitably modified. Let σ0 be a given polyhedral k-current, for ρε
a standard approximation of the identity we let σ be a a smooth k-current such that ∂σ = ∂σ0∗ ρε.
(In equation (19) the current is identified with its density measure.) In the chapter we introduce formally the energy and show that it Γ-converges to the energy Eh defined
+ + + + − − − −
Figure 6: Computed mass flux σ and phase fields ϕ1, ϕ2, ϕ3 for the cost function shown
on the right, ε = 0.005. The color in σ indicates which phase field is active. The result is obtained by optimizing the functional defined in Chapter Three.
In the Fourth and Fifth Chapters we restrict again our attention to sets Ω⊂ R2 and develop two functionals for the approximation of any concave and continuous transport cost function h. Note that we say that a transport cost function is concave if it is an even function whose restriction to [0, +∞) is concave. The first result regards transport cost functions h of the form
h(m) = min{αi|m| + βi : 0≤ i ≤ N}.
for α0 > α1 > . . . > αN ≥ 0 and 0 ≤ β0 < β1 < . . . < βN. Our approach takes
advantage of the result in the First Chapter in which we recovered in the Γ-limit affine cost functions of the form 1 + β|m|. In the case N = 1 and β0 > 0 the proposed
phase-field energy takes the form Z Ω min ϕ20+α 2 0ε2 β0 ; ϕ21+α 2 1ε2 β1 |σ|2 2ε dx + β0Tε(ϕ0) + β1Tε(ϕ1)
where Tε is an energy of the Modica-Mortola type defined as
Tε(ϕ) = 1 2 Z Ω ε|∇ϕ(x)|2+(ϕ(x)− 1) 2 ε dx.
Let us highlight the presence of two phase-fields which interact in the constraint compo-nent of the functional. Ideally each 1−ϕi is a smooth indicator function of some subset
of the support of the limit rectifiable measure σ. In particular ϕi = 0 if the choice of
the i-th component in the definition of h is optimal with respect to the intensity of the flux of σ. The entire Fourth Chapter is devoted to establish the proof of the Γ-convergence result and the study of numerical methods developed in collaboration with Carolin Rossmanith and Benedikt Wirth from Munster University.
In the final chapter of the thesis we study functionals of the form
Fε(σ, ϕ) := Z Ω f (ϕ)|σ| + 1 2 ε|∇ϕ|2+ϕ 2 ε dx
The two main differences with respect to the previous models are the linear penalization in|σ| and the presence of the term ϕ2instead of (1−ϕ)2. Analogous models with a linear
penalization of the |σ| component have been studied recently in the case of fracture theory and the generalized Mumford-Shah functional [ABS99, DMOT16]. Our main
contribution is to find an explicit form of the weight function f to obtain in the limit the energyEh. For a continuous and concave transport cost function h, we define f as
f (t) = (−h∗)−1(t2).
The function h∗ is the (concave) Legendre transform of h. In this model ϕ takes value
0 and not 1 outside the support of the limit measure σ. By virtue of this general result we address the problem of the numerical approximation of the functional Fε.
The linear penalization in σ may be seen as a drawback with respect to the methods previously studied which where deeply based on the quadratic cost |σ|2. In view of this difference we started investigating new numerical methods based on the Beckman model [Bec52] for transportation. The same result may be obtained with different choice of the well potential. Namely, given a potential W which is an even function, increasing on [0, +∞) and vanishing in 0 we introduce the transition energy
cW(t) :=
Z |t|
0
2pW (s) ds.
Then choosing f (t) = (−h∗)−1◦ cW(t) the same Γ-convergence result may be obtained
with a family of functionals defined as
Fε(σ, ϕ) := Z Ω f (ϕ)|σ| + 1 2 ε|∇ϕ|2+W (ϕ) ε dx.
In force of this degree of freedom in the choice of the potential W we start analyzing which would be the best choice. These and other questions are the subject of the concluding section which investigates possible developments of the proposed methods.
Affine cost function
1.1
Introduction
In this chapter we devise an approximation for the minimization problem defined in the introduction in equation (6). In particular we consider the energy Eh choosing as
cost function
h(m) := (
1 + β|m|, if m6= 0 0, otherwise,
where β > 0 is a fixed positive parameter. Furthermore we will consider only atomic probability measures µ+, µ− ∈ P(Ω) to define the constraint. The results contained in
this chapter have been published in the paper [CFM17a]. Let us introduce the precise framework of our approximation. Let ρ ∈ Cc∞(R2, R
+) be a classical radial mollifier
with supp ρ ⊂ B1(0) and R ρ = 1. For ε ∈ (0, 1], we set ρε(x) = ε−2ρ(ε−1x) and we
define the space Vε(Ω) of square integrable vector fields with weak divergence satisfying
the constraint
∇ · σε = (µ+− µ−)∗ ρε in D0(R2). (1.1)
For an η = η(ε) > 0, we denote
Wε(Ω) = ϕ ∈ W1,2(Ω) : η ≤ ϕ ≤ 1 in Ω, ϕ ≡ 1 on ∂Ω .
We denote with Xε(Ω) = Vε(Ω) × Wε(Ω) and define the energy Fε : M(Ω, R2) ×
L1(Ω)→ [0, +∞] as Fε(σ, ϕ) := Z Ω 1 2εϕ 2 |σ|2 dx + Z Ω ε 2|∇ϕ| 2+(1− ϕ)2 2ε dx, if (σ, ϕ)∈ Xε(Ω), +∞, otherwise. (1.2)
From now on, we assume that
η ε
ε↓0
−→ β. (1.3)
We denote MS(Ω) the set of R2-valued measures σ ∈ M(R2, R2) with support in Ω
such that the constraint
holds. We define the limit energy Eβ :M(Ω, R2)× L1(Ω)→ [0, +∞] as Eβ(σ, ϕ) = Z Σ (1 + β m) dH1 if ϕ≡ 1, σ ∈ MS(Ω) and σ = (m, τ, Σ), +∞ otherwise. (1.5)
We prove the Γ-convergence of the sequence (Fε) to the energy Eβ as ε ↓ 0. More
precisely the convergence holds in M(Ω, R2)× L1(Ω) where M(Ω, R2) is endowed
with the weak-∗ topology and L1(Ω) is endowed with its classical strong topology. We begin by proving the equicoercivity of the sequence (Fε). In this statement and
throughout the chapter, we make a small abuse of language by denoting (aε)ε∈(0,1] and
calling sequence a family {aε} labeled by a continuous parameter ε ∈ (0, 1]. In the
same spirit, we call subsequence of (aε), any sequence (aεj) with εj → 0 as j → +∞.
We establish the following lower bound.
Theorem 1.1 (Γ− lim inf). For any sequence (σε, ϕε)⊂ M(Ω, R2)× L1(Ω) such that
σε ∗
* σ and ϕε→ ϕ in the L1(Ω) topology, with (σ, ϕ)∈ M(Ω, R2)× L1(Ω),
lim inf
k→+∞ Fε(σε, ϕε)≥ Eβ(σ, ϕ).
To complete the Γ-convergence analysis, we establish the matching Γ-limsup in-equality.
Theorem 1.2 (Γ−lim sup). For any (σ, ϕ) ⊂ M(Ω, R2)×L1(Ω) there exists a sequence
(σε, ϕε) such that σε ∗
* σ and ϕε → ϕ in the L1(Ω) topology and
lim sup
k→+∞
Fε(σε, ϕε)≤ Eβ(σ, ϕ).
Theorem 1.3 (Equicoercivity). Assume β > 0. For any sequence (σε, ϕε)ε∈(0,1] ⊂
M(Ω, R2)× L1(Ω) with uniformly bounded energies, i.e.
sup
ε Fε
(σε, ϕε) < +∞,
there exist a subsequence εj ↓ 0 and a measure σ ∈ MS(Ω, R2) such that σεj → σ with
respect to the weak-∗ convergence of measures and ϕεj → 1 in L
1(Ω). Moreover, σ is
a rectifiable measure (i.e., it is of the form σ = (m, τ, Σ)).
Structure of the chapter: In Section 1.2 we introduce and recall some notation and several tools and notions on SBV functions and introduce an operator acting on vector field measures. In Section 1.3 we study the behavior of the functional Fε on
vector fields of the form ∇u (dropping the divergence constraint). In Section 1.4 we prove the equicoercivity result, Theorem 1.3 and we establish the lower bound stated in Theorem 1.1. In Section 1.5 we prove the upper bound of Theorem 1.2. Finally, in the last section, we present and discuss various numerical simulations.
1.2
Preliminaries for the chapter
In the following Ω ⊂⊂ ˆΩ ⊂ Rd are bounded open convex sets. Given X ⊂ Rd (in
practice X = Ω or X = ˆΩ), we denote by A(X) the class of all relatively open subsets of X and by AS(X) the subclass of all simply connected relatively open sets O ⊂ X
such that O∩ S = ∅. We denote by (e1, . . . , ed) the canonical orthonormal basis of
Rd, by | · | the euclidean norm and by h·, ·i the euclidean scalar product in Rd. The open ball of radius r centered at x∈ Rd is denoted by B
r(x). The (d− 1)-dimensional
Hausdorff measure in Rd is denoted by Hd−1. We write |E| to denote the Lebesgue
measure of a measurable set E ⊂ Rd. When µ is a Borel meaure and E ⊂ Rd is a Borel set, we denote by µ
x
E the measure defined as µx
E(F ) = µ(E∩ F ).Let us remark that from Section 1.4 onwards, we work in dimension d = 2.
For any fixed couple (σ, ϕ), withFε(σ, ϕ; O) we denote the value of the functional (1.2)
on any set O∈ A(Ω). Similarly we define the with version Eβ(σ, ϕ; O) the localization
of Eβ to O.
BV(Ω) is the space of functions u ∈ L1(Ω) having as distributional derivative Du
a measure with finite total variation. Following the classical notation as in [AFP00, ABM14] and [Bra98] for u∈ BV (Ω) we have
Du =∇u dx + (u+− u−)νuHd−1
x
Ju+ Dcu,where Ju is the set of “approximate jump points” x where y 7→ u(x + ρy) converge as
ρ→ 0 to u+χ{y·νu≥0}+ u
−χ
{y·νu<0}for some (u
−, u+, ν
u) and Dcu is the Cantor “part”.
Let us introduce the space of special functions of bounded variation and a variant:
SBV (Ω) :={u ∈ BV (Ω) : Dcu = 0},
GSBV (Ω) :={u ∈ L1(Ω) : max(−T, min(u, T )) ∈ SBV (Ω) ∀ T > 0}.
Eventually, in Section 1.3, the following space of piecewise constant functions will be useful.
P C(Ω) ={u ∈ GSBV (Ω) : ∇u = 0}. (1.6) To conclude this section we recall the slicing method for functions of bounded variation. Let τ ∈ Sd−1 and let
Πτ :={y ∈ Rd:hy, τi = 0}.
If y∈ Πτ and E ⊂ Rd, we define the one dimensional slice
Eτ,y :={t ∈ R : y + tτ ∈ E}.
For u : Ω→ R, we define uτ,y : Ωτ,y → R as
uτ,y(t) := u(y + tτ ), t ∈ Ωτ,y.
Functions in GSBV (Ω) can be characterized by one-dimensional slices (see [Bra98, Thm. 4.1])
Theorem 1.4. Let u∈ GSBV (Ω). Then for all τ ∈ Sd−1 we have
uτ,y ∈ GSBV (Ωτ,y) for Hd−1-a.e. y∈ Πτ.
Moreover for such y, we have
u0τ,y(t) =h∇u(y + tτ), τi for a.e. t ∈ Ωτ,y,
Juτ,y ={t ∈ R : y + tτ ∈ Ju},
and
uτ,y(t±) = u±(y + tτ ) or uτ,y(t±) = u∓(y + tτ )
according to whether hνu, τi > 0 or hνu, τi < 0. Finally, for every Borel function
g : Ω→ R, Z Πτ X t∈Juτ,y gτ,y(t) dHd−1(y) = Z Ju g|hνu, τi| dHd−1. (1.7)
Conversely if u ∈ L1(Ω) and if for all τ ∈ {e1, . . . , ed} and almost every y ∈ Πτ we
have uτ,y ∈ SBV (Ωτ,y) and
Z
Πτ
|Duτ,y|(Ωτ,y) dHd−1(y) < +∞
then u∈ SBV (Ω).
Let us introduce the linear operator ⊥ that associates to each vector v = (v1, v2)∈
R2 the vector v⊥= (
−v2, v1) obtained via a 90◦ counterclockwise rotation of v. Notice
that the⊥ operator maps divergence-free R2-valued measures onto curl free R2-valued measures. Let O ⊂ R2 be a simply connected and bounded open set. It is possible
to generalize Stokes Theorem to divergence free measures. If µ is a smooth divergence free vector field on O we have µ = ∇u⊥ for some smooth function with zero mean value. Then by Poincar´e inequality |u|L1 ≤ C|µ|L1. The result for µ general divergence
free finite vector measure follows by regularization. On the other hand for u ∈ P C(Ω), σ := Du⊥ is divergence free and,
σ = (u+− u−)νu⊥H1 = U (Ju, [u], νu⊥). (1.8)
1.3
Local Result
In this section we introduce a localization of the family of functionals (Fε) (see (1.2)).
We establish a lower bound and a compactness property for these local energies. In this section we assume Ω⊂ Rd.
Localization. Let O∈ AS(Ω) be a simply connected relatively open subset of Ω. For
uε ∈ W1,2(O) and ϕε∈ W1,2(O), we define
i.e. as the evaluation of the functional Fε on vector fields of the form ∇u with no
requirement on the divergence. Notice that for ε < d(O, S), we have ∇ · σε ≡ 0 in O
for any σε ∈ Vε(Ω). By Stokes theorem we have Du⊥ε = σε for some uε ∈ W1,2(O) and
we have
Fε(σε, ϕε; O) =Gε(uε, ϕε; O).
The rest of the section is devoted to the proof of
Theorem 1.5. Let (uε)ε∈(0,1] ⊂ W1,2(O) be a family of functions with zero mean
value and let (ϕε) ⊂ W1,2(O) such that ϕε ∈ W1,2(O, [η(ε), 1]). Assume that c0 :=
supεGε(uε, ϕε; O) is finite. Then there exist a subsequence εj and a function u∈ BV (Ω)
such that
a) ϕεj → 1 in L
2(O),
b) uεj → u with respect to the weak-∗ convergence in BV ,
c) u∈ P C(O).
Furthermore for any piecewise function u ∈ P C(O) and any sequence (uε, ϕε) such
that uε ∗
* u and ϕε → 1, we have the following lower bound of the energy:
lim inf
ε→0 Gε(uε, ϕε; O)≥
Z
Ju∩O
(1 + β|[u]|) dHd−1.
The proof is achieved in several steps and mostly follows ideas from [Iur13] (see also [CFI16]). In the first step we obtain (a) and (b). In step 2 we prove (c) and the lower bound for one dimensional slices of Gε. Finally in step 3 we prove (c) and
the lower bound in dimension d. The construction of a recovery sequence that would complete the Γ-limit analysis is postponed to the global model in Section 1.5.
Proof. Step 1: Item (a) is a straightforward consequence of the definition of the func-tional. Indeed, we have
Z
O
(1− ϕε)2 dx ≤ 2 ε Gε(uε, ϕε) ≤ 2, c0ε ε↓0
−→ 0.
For (b), since (uε) has zero mean value, we only need to show that supε∈(0,1] |Duε|(O) <
+∞. Using Cauchy-Schwarz inequality we get [|Duε|(O)] 2 = Z O |∇uε| 2 ≤ 2 ε Z O 1 ϕ2 ε 1 ε Z O ϕ2ε|∇uε|2 . (1.9)
By assumption, the second therm in the right hand side of (1.9) is bounded by 2c0. In
order to estimate the first term we split O in the two sets{ϕε< 1/2} and {ϕε ≥ 1/2}.
We have, 2 ε Z O 1 ϕ2 ε = 2 ε Z {ϕε<1/2} 1 ϕ2 ε + Z {ϕε≥1/2} 1 ϕ2 ε .
Since η≤ ϕε≤ 1/2 on {ϕε ≤ 1/2} it holds ϕ2ε(1− ϕε)2 ≥ η2(1− 1/2)2 therefore Z {ϕε<1/2} 1 ϕ2 ε ≤ 2ε η2(1− 1/2)2 Z {ϕε<1/2} (1− ϕε)2 2ε ≤ 8ε η2c0, Z {ϕε≥1/2} 1 ϕ2 ε ≤ Z {ϕε≥1/2} 1 (1/2)2 = 4|{ϕε≥ 1/2}|.
Eventually, as |{ϕε≥ 1/2}| ≤ |O|, combining these estimates with (1.9) we obtain
[|Duε|(O)] 2 ≤ ε 2 η2 16c 2 0+ 8ε|O|c0 ε↓0 −→ 16c 2 0 β2 < ∞. (1.10) This establishes (b).
Step 2: In this step we suppose O to be an interval of R, so that uε, ϕε are
one-dimensional. We first prove that u is piecewise constant. The idea is that in view of the constraint component of the energy, variations of uε are balanced by low values
of ϕε. On the other hand the Modica-Mortola component of the energy implies that
ϕε ' 1 in most of the domain and that transitions from ϕε ' 1 to ϕε ' 0 have a
constant positive cost (and therefore can occur only finitely many times). Step 2.1: (Proof of u∈ P C(O).) Let us define
Bε:= x∈ O : ϕε(x) < 3 4 ⊃ Aε := x∈ O : ϕε(x) < 1 2 , (1.11) and let Cε ={I connected component of Bε : I∩ Aε 6= ∅}. (1.12)
Let us show that the cardinality of Cε is bounded by a constant independent of ε. Let
ε be fixed and consider an interval I ∈ Cε. Let a, b ∈ ¯I such that {ϕε(a), ϕε(b)} =
{1/2, 3/4}. Using the usual Modica-Mortola trick, we have Gε(uε, ϕε; I) ≥ Z I ε|ϕ0ε|2 2 + (1− ϕε)2 2ε dx ≥ Z (a,b) |ϕ0 ε|(1 − ϕε) dx ≥ Z 3/4 1/2 (1− t) dt = 3 25.
Since all the elements of Cε are disjoint and Gε(uε, ϕε,·) is additive, we deduce from
the energy bound that
#Cε≤ 25c0/3,
where we denote #Cε the cardinality of Cε. Next, up to extracting a subsequence we
assume that #Cε = N is fixed. The elements of Cεare of the form Iiε= (mεi− wεi, mεi+
wεi) for i = 1,· · · , N, with mεi < mεi+1. Since ϕε → 1 in L1(O) we have
X Iε i∈Cε |Iε i| = N X i=1 2wiε→ 0. (1.13)
Up to extracting a subsequence, we can assume that each sequence (mε
i) converges in
O. We call m1 ≤ m2 ≤ · · · ≤ mN their respective limits. We now prove that
|Du|(O \ {mi}Ni=1) = 0, (1.14)
thus supp(|Du|) ⊂ {m1,· · · , mN}. The latter ensures that u has no Cantor component
since Du is supported on a finite number of points and that is a.e. constant outside {m1,· · · , mN} so that u ∈ P C(O), (1.6). To this aim, we fix x ∈ O \ {mi}Ni=0 and
establish the existence of a neighborhood Bδ(x) of x for which |Du|(Bδ(x)) = 0. Let
0 < δ≤ mini|x−mi|/2. Equation (1.13) ensures that for ε small enough Bδ(x)∩Cε=∅.
Notice that from the definitions in (1.11) and (1.12) we have that ϕε ≥ 1/2 outside the
union of the sets in Cε. Hence, using Cauchy-Schwarz inequality, we have for ε small
enough, Z Bδ(x) |u0ε| dx 2 ≤ 2δ Z Bδ(x) |u0ε| 2 dx ≤ (2δ)(2ε)4 1 2ε Z Bδ(x) ϕ2ε|u0ε|2 dx ≤ 16c0εδ ε↓0 −→ 0.
By lower semicontinuity of the total variation on open sets we conclude that|Du|(Bδ(x)) =
0, which proves the claim (1.14).
Step 2.2: (Proof of the lower bound for Gε.) Without loss of generality we can
assume N = 1, thus Ju is composed of a single point, otherwise the argument we
propose can be applied on each mi separately. Up to a translation m1 = 0 and we
denote D := u(0+) = −u(0−) > 0. For any 0 < d < D there exist six points
y1 < x1ε ≤ ˜x1ε < ˜x2ε ≤ x2ε < y2 such that lim ε→0ϕε(y1) = limε→0ϕε(y2) = 1, lim ε→0ϕε(x 1 ε) = limε→0ϕε(x2ε) = 0, (1.15) uε(˜x1ε) = −D + d, uε(˜x2ε) = D− d.
Since ϕε → ϕ and uε → u in L1 up to a subsequence they converge point-wise almost
everywhere and this implies the first and third fact. Let inf(y1,y2)ϕε = cε, then Jensen
inequality implies c0 ≥ Z y2 y1 ϕ2 ε|u 0 ε|2 2ε dx≥ c2 ε 2ε(y2− y1) Z y2 y1 |u0ε| dx 2 .
Then cεmust vanish with ε implying statement (1.15). Using the Modica-Mortola trick
in the intervals (y1, x1ε) and (x2ε, y2) as above, we compute:
lim inf ε↓0 Gε(uε, ϕε; (y1, x 1 ε)∪ (x 2 ε, y2))≥ ≥ lim inf ε↓0 Z x1ε y1 (1− ϕε)|ϕ0ε| dx + Z y2 x2 ε (1− ϕε)|ϕ0ε| dx ! ≥ 1. (1.16)
For the estimate on the interval Iε = (˜x1ε, ˜x2ε) let us introduce: Gε :=w ∈ W1,2(Iε) : w(˜x1ε) = −D + d, w(˜x 2 ε) = D− d , Zε :=z ∈ W1,2(Iε) : η ≤ z ≤ 1 a.e. on Iε , Hε(w, z) := Z Iε 1 2εz 2 |w0|2+ (1− z)2 2ε dx, hε(z) = inf w∈Gε Hε(w, z) for z ∈ Zε.
Note that for w ∈ Gε and z ∈ Zε, we can apply an inequality similar to (1.9). In
particular, for z replacing ϕε and w0 taking the place of Duε we get
Z Iε |w0 | dx 2 ≤ Z Iε z2|w0|2 Z Iε 1 z2 .
Reversing the latter and taking into account the conditions on w obtains
Z Iε z2|w0|2 ≥ Z Iε |w0| dx 2Z Iε 1 z2 −1 ≥ 4(D − d)2 Z Iε 1 z2 −1 .
From this we deduce the lower bound
hε(z) ≥ 4(D − d)2 2ε Z Iε 1 z2 dx −1 + Z Iε (1− z)2 2ε dx. (1.17) Let us remark that optimizing Hε(w, z) with respect to w ∈ Gε we see that this
inequality is actually an equality. Consider for 0 < λ < 1 the inequalities: Z {x∈Iε:ϕε≥λ} 1 ϕ2 ε ≤ |Iε| λ2 , Z {x∈Iε:ϕε<λ} 1 ϕ2 ε ≤ 1 (1− λ)2 2ε η2 Z Iε (1− ϕε)2 2ε dx .
Applying both of them in (1.17) we obtain Gε(uε, ϕε, Iε)≥ hε(ϕε) ≥ 2(D− d) 2 ε|Iε| λ2 + 1 (1−λ)2 2ε2 η2 R Iε (1−ϕε)2 2ε dx + Z Iε (1− ϕε)2 2ε dx ≥ 2(1 − λ)η ε(D− d) − (1 − λ) 2η 2 2ε |Iε| λ2 (1.18)
where the latter inequality is obtained by minimizing the function:
t7→ 2(D− d) 2 ε|Iε| λ2 + 1 (1−λ)2 2ε2 η2 t + t.
Therefore we can pass to the limit in (1.18) and obtain:
lim inf
Sending λ and d to 0 and recalling the estimate in (1.16) we get lim inf
ε↓0 Gε(uε, ϕε, (y1, y2)) ≥ 1 + β 2D = 1 + β|u(0 +)
− u(0−)|. (1.19) Step 3: Indeed by Fatou’s Lemma for any τ ∈ Sd−1 and Hd−1 almost every y ∈ Ω
τ it holds lim inf ε↓0 Gε(uε, ϕε; O) ≥ Z Πτ lim inf ε↓0 Z Oτ y 1 2ε(ϕ 2 ε) τ y|(u 0 ε) τ y| 2 + ε 2|(ϕ 0 ε) τ y| 2 + (1− (ϕε) τ y)2 2ε dt dH d−1 (y).
Then by the results in Step 2.1 and 2.2, in particular inequality (1.19), it holds
lim inf ε↓0 Gε(uε, ϕε; O)≥ Z Πτ X mi∈(Ju)τy 1 + β|uτ y(m + i )− u τ y(m − i )| dH d−1 (y).
Therefore by Theorem 1.4 we have u ∈ SBV (O). Moreover, since (u0)τ
y = 0 on each
slice, we have u∈ P C(O). Applying identity (1.7) we get lim inf
ε→0 Gε(uε, ϕε; O)≥
Z
Ju∩O
|νu· τ| [1 + β|[u]|] dHd−1. (1.20)
In order to conclude, we use the following localization method stated by Braides in [Bra98, Prop. 1.16].
Lemma 1.1. Let µ :A(X) → [0, +∞) be an open-set function superadditive on open sets with disjoint compact closures and let λ be a positive measure on X. For any i∈ N let ψi be a Borel function on X such that µ(A)≥
R
Aψi dλ for all A∈ A(X). Then
µ(A) ≥ Z
A
ψ dλ
where ψ := supiψi.
For any u∈ P C(O) let us introduce the increasing set function µ defined on A(O) by µ(A) := inf (ϕε,uε)→(1,u) n lim inf ε→0 Gε(uε, ϕε; A) o
, for any A ∈ A(O).
Observe that for any two open sets A and B with disjoint compact closure and for any (uε, ϕε) such that uε
∗
* u and ϕε → 1 on A ∪ B, the restriction of uε to A (resp. B)
weak-∗ converges in A (resp. B) to the restriction of u on A (resp. B) and it follows µ(A + B) ≥ µ(A) + µ(B).
This proves that µ is superaddittive on open sets with disjoint compact closures. Let λ be a Radon measure defined as
Fix a sequence (τi)i∈N dense in Sd−1. By (1.20) we have µ(O)≥ Z O ψi dλ, i∈ N, where ψi(x) := ( |hνu(x), τii|, if x ∈ Ju, 0, if x∈ O \ Ju.
Hence by Lemma 1.1 we finally obtain
lim inf ε→0 Gε(uε, ϕε; O)≥ Z O sup i ψi(x) dµ = Z Ju∩O [1 + β|[u]|] dHd−1.
1.4
Equicoercivity and Γ-liminf
From now till the end of the chapter we assume that Ω ⊂ R2. Let us first produce the
following construction.
Lemma 1.2. Given two probability measures µ+ and µ− supported on a finite set of
points S = {x0, . . . , xN}, there exists a vector measure γ = U(mγ, τγ, Σγ) and a finite
partition (Ωi)⊂ A(Ω) of Ω such that
a) ∇ · γ = −µ++ µ−,
b) each Ωi is a polyhedron,
c) Σγ ⊂Si∂Ωi,
d) Ωi is of finite perimeter for each i and Ωi∩ Ωj =∅ for i 6= j,
e) |Ω \ ∪iΩi| = 0.
Moreover if M is a 1 dimensional countably rectifiable set, we can choose γ and (Ωi)
such that H1(M ∩S
i∂Ωi) = 0.
Proof. Let us fix a point p∈ Ω \ S and assume µ+− µ− =
N
X
i=0
aiδxi.
Consider the map xi−p
|xi−p|t + p : [0, 1]→ Ω, then the measure
γi = p− xi |p − xi|· +x i # [0, 1]
is supported on the segment [p, xi] =: Σγ and is such that ∇ · γi = δxi − δp for i ∈
{0, · · · , N}. We define γ = N X i=1 aiγi.
By construction (a) holds true. Moreover, up to a small shift of p we may assume that [p, xi]∩ [p, xj] ={p} for i 6= j.
Next, let Dj be the straight line supporting [p, xj]. We define the sets (Ωi) as the
connected components of Ω\ (D0∪ · · · ∪ DN). We see that (c, d, e) hold true.
For the last statement, we observe that by the coarea formula, we haveH1(Σ∩S
i∂Ωi) =
0 for a.e. choice of p.
Figure 1.1: Example of the construction of the H1-rectifiable measure γ (red) and of the partition {Ωi} (gray) in the case M (green) is being a H1-rectifiable set. Here
µ+ = δx0 and µ−= 1/3(δx1 + δx2 + δx3).
We now prove the compactness property (Theorem 1.3). Let us consider a sequence (σε, ϕε)∈ M(Ω, R2) uniformly bounded in energy by c0 < +∞,
0≤ Fε(σε, ϕε)≤ c0 for ε∈ (0, 1]. (1.21)
Proof of Theorem 1.3. First observe that by definition (1.2) and equation (1.21), we have σε ∈ Vε(Ω) and ϕε ∈ Wε(Ω).
Next, using the arguments of Step 1 of the proof of Theorem 1.5, with |σε| instead
of |∇uε|, inequality (1.10) reads
|σε|(Ω) ≤ s 16ε 2 η2 c 2 0+ 8ε|Ω|c0 ε↓0 −→ 4cβ0 < ∞. (1.22) Thus the total variation of (σε)εis uniformly bounded as long as β > 0 and there exists
a σ∈ MS(Ω) such that up to a subsequence σε ∗
* σ inM(Ω). Now, considering the last term in the energy (1.2) we have
Z
Ω
(1− ϕε)2 dx≤ 2ε Fε(σε, ϕε)≤ 2ε c0 → 0.
Hence, ϕε → 1 in L2(Ω).
Let us now study the structure of the limit measure σ. Let us recall that ˆΩ is a bounded convex relatively open set such that Ω ⊂ ˆΩ and let us extend σε by 0 and
ϕε by 1 in ˆΩ\ Ω. Obviously we have Fε(σε, ϕε; ˆΩ) = Fε(σε, ϕε; Ω), therefore for any
O ∈ AS( ˆΩ) applying the localization described in Section 1.3 we can associate to each
σε a function uε ∈ W1,2(O) with mean value 0 such that σε = ∇u⊥ε in O. Since
|∇u⊥
ε| = |∇uε| by Theorem 1.5 there exists a u ∈ P C(O) such that, up to extracting
a subsequence, uε ∗
* u. Eventually, from formula (1.8), we get
σ
x
O = Du⊥x
O =−[u]νJ⊥uH1x
(Ju∩ O).Since we can cover Ω \ S by countable many sets O ∈ AS( ˆΩ), this shows that σ
decomposes as
σ = (mσ, τσ, Σσ) + ω,
where ω is a measure absolutely continuous with respect to H0
x
S. By Lemma 1.2there exists a rectifiable measure γ = U (mγ, τγ, Σγ) such that ∇ · (σ + γ) = 0 and
H1(Σ
γ ∩ Σσ) = 0. Then there exists a u ∈ BV (Ω) such that Du = σ⊥+ γ⊥. Since
u ∈ BV (Ω) and S is composed by a finite number of points, we deduce |Du|(S) = 0 which implies |ω|(S) = 0. Hence σ writes in the form (mσ, τσ, Σσ).
Let us now use the local results of Section 1.3 to prove the Γ− lim inf inequality. Proof of Theorem 1.1. Let (σε, ϕε) such that σε
∗
* σ and ϕε → ϕ as in the statement
of the theorem. Without loss of generality, we can suppose that Fε(σε, ϕε) < +∞.
Let ˆΩ be as in the proof of Theorem 1.3 and let us define χ = Γ− lim infεFε(σε, ϕε)
and λ = β|σ| + H1
x
Σσ. Consider the countable family of sets {Oi} ⊂ AS( ˆΩ) made ofthe relatively open rectangles Oi ⊂ ˆΩ\ S with vertices in Q2. The local result stated
in Theorem 1.5 gives for any i∈ N
χ(A)≥ µ(Oi∩ A) ≥ λ(Oi∩ A) =
Z
A
ψi dλ,
where ψi := 1Oi. Therefore Lemma 1.1 gives
Γ− lim inf
ε↓0 Fε(σε, ϕε) = µ( ˆΩ)≥ λ(ˆΩ) = β|σ|(Ω) + H 1(Σ
σ)
since supiψi is the constant function 1.
1.5
Γ-limsup inequality
Let us prove the Γ-limsup inequality stated in Theorem 1.2. Recall that the latter consists in finding a sequence (σε, ϕε) for any given couple (σ, ϕ)∈ M(Ω, R2)× L1(Ω)
such that σε ∗ * σ, ϕε → ϕ in L1(Ω) and lim sup ε↓0 F ε(σε, ϕε)≤ Eβ(σ, ϕ). (1.23)
When Eβ(σ, ϕ) = +∞ the inequality is valid for any sequence therefore by
White [Whi99b], [Whi99a] and Xia [Xia03] polyhedral vector measures are dense in energy and it is sufficient to consider vector measures of the form
σ =
n
X
i=1
(mi, τi, Σi), (1.24)
where Σi is a segment, mi ∈ R+ is H1-a.e. constant and τi is an orientation of Σi
for each i. We included in appendix A a proof of this result based on BV functions. Without loss of generality we can suppose that for each couple of segments Σi, Mj,
for i 6= j, the intersection Σi ∩ Mj is at most a point (called branching point) not
belonging to the relative interior of Σi and Mj. We first produce the estimate (1.23)
for σ concentrated on a single segment thus let us assume σ = me1H1
x
(0, l)× {0}.Notation: Let us fix the values
aε:= mβ ε 2 if β > 0 ε if β = 0 , bε := ε ln 1− η ε and rε = max{ε, aε}.
Let d∞(x, S) be the distance function from x to the set S ⊂ Ω relative to the infinity
norm on R2 and Qr(P ) ={x ∈ R2 : d∞(x, P )≤ r} the square centered in P of size 2r
and sides parallel to the axes. Introduce the sets Iaε :={x ∈ R 2 : d ∞(x, [0, l]× {0}) ≤ aε} ∪ Qrε(0, 0)∪ Qrε(l, 0), Ibε :={x ∈ R 2 : d ∞(x, Iaε)≤ bε} \ Iaε, Icε :={x ∈ R 2 : d ∞(x, (Iaε ∪ Ibε))≤ ε} \ (Iaε ∪ Ibε) , Idε := Ω\ (Iaε ∪ Ibε ∪ Icε), Σε(t) :={(t, x2) : |x2| ≤ rε}, and define Rε= Iaε\ (Qrε(0, 0)∪ Qrε(l, 0)). l l Bε Bε Bε Bε Σε(rε) Σε(l− rε) Σε(rε) Σε(l− rε) Rε Rε Qrε(0, 0) Qrε(l, 0) Qrε(0, 0) Qrε(l, 0) Ibε Icε Ibε Icε
Figure 1.2: Example of the neighborhoods of the segment [0, l]× {0}. On the left the case in which rε = ε, on the right the case in which rε= aε > ε. The cyan region is Rε
and Iaε = Rε∪ (Qrε(0, 0)∪ Qrε(l, 0)). Remark that supp(ρε) = B(0, ε).
Costruction of σε: We build σε as a vector field supported on Iaε. In particular we
add together three different constructions performed respectively on Rε, Qrε(0, 0) and
Qrε(l, 0). We construct σε on Rε in order to obtain the Γ-limsup inequality, on the